One of the best parts of my job is the exposure to diversity of businesses. In one recent week I spoke with the CIO of an electronics manufacturer, the head of research of a cosmetics manufacturer, and the chief scientist of a metals mining company. What these leaders have in common is a desire to use artificial intelligence (AI) to optimize their businesses, and all of them had an idea as to where AI might apply for their business. Interestingly, though, these three conversations converged on the same topic: the need to specify concrete outcomes for your AI to work toward.
It may seem counterintuitive that artificial intelligence, a system whose entire purpose is to learn on its own, needs to be told what to do. Nonetheless, it’s true: in order for AI to produce value for a business, the business must first determine the specific outcomes for which they want to apply AI.
Take the example of customer attrition. Attrition prediction (and its sibling, customer retention) is one of the most common AI use cases for any business. Nobody wants to lose a customer. But what is attrition exactly? What does it mean for your business to lose a customer? Consider the case of customer attrition in a bank. If a bank’s customer holds a checking account and a savings account with a bank, and that customer closes the savings account, is that attrition? What if that customer has just a single account with the bank but she just moved 90% of the assets out if it – is that attrition?
Before you can use AI to help predict and reduce your customer attrition, you must first come up with a concrete notion of what attrition is for your business. That brings me to the first truism of artificial intelligence:
If you can’t report, you can’t predict.
The often overlooked first step in implementing an AI system has nothing to do with AI itself. It’s all about historical reporting. To continue the bank customer attrition example above, if the bank in question wants to use AI to predict their attrition, they’re going to need to define attrition very concretely, and the way one usually does so is by reporting on historical attrition. The bank would want to build a report that shows who has attrited according to their definition of attrition. For each customer it should come down to a yes and no question. Yes or no: did the customer who withdrew 90% of her assets attrit? Yes or no: did the customer who just closed the savings account attrit? Sometimes this forces the business to divide the question into parts – maybe there needs to be two different measures of attrition, one for checking accounts and one for savings accounts.
As simple as this may sound, it is often quite complex. Many times a business has only a fuzzy notion of the outcome, like attrition, that it wants to optimize. Sometimes the data that would be used to build such a report is spread across many disparate systems, forcing the business users to create integrations to get all the data in one place. It can be painful, but it is necessary: you’ll need to take these steps anyway in order to get AI to work for you.
The above truism lends itself to the following corollary:
If you do report, you often want to predict.
Most businesses today are analytics-driven, and already run many in-house reports to measure the health of their business. The outcomes that they’re measuring in these reports are often the low-hanging fruit for AI optimization. The outcomes have already been decided; the data is already there.
In the case of the cosmetics research department I spoke to, they suspected that certain combinations of ingredients and product packaging were causing quality issues that they wanted to root out. This is a slam-dunk case for AI, because this company runs their business on product quality. They report on quality issues extensively, and have been gathering data about which ingredients and packaging they’ve been using since the first bottle of moisturizing cream rolled off the line. That makes predicting product quality a fantastically doable scenario for this business.
In sum, to make the most of AI, you must concretely define the outcomes you’re trying to optimize, and the best way to do so is by reporting on those outcomes historically. And if there are certain reports on which you run your business, look there first – that’s going to be the best place to start your journey with AI.